ai-digest.dev
last updated 3 h ago
CodingarXiv cs.AI 18 d ago

CodeTeam: An LLM-Powered Multi-Agent Framework for Repository-Level Code Generation

CodeTeam is an LLM-based multi-agent framework designed for natural language to repository generation (NL2Repo), which separates the software development process into planning, decision-making, and implementation stages. It employs multiple Architect agents for drafting software design sketches, a CTO agent for evaluation and normalization, and Developer agents for code generation, achieving significant improvements on the SketchEval benchmark with an increase of 4.1 and 2.9 points in SketchBLEU for prompt-engineering and supervised fine-tuning variants, respectively. The framework also demonstrates a high average test pass rate of 34.6% and 42.3% on the NL2Repo-Bench benchmark, highlighting its effectiveness in producing functional code and its relevance for practitioners in automating repository-level code generation.

code generationmulti-agentllmrelevance 0.00 · engagement 0.00
Read at source ↗← all news
CodeTeam: An LLM-Powered Multi-Agent Framework for Repository-Level Code Generation — AI News Digest